China's new frontier in policing is the body itself
On 11 June 2026, the South China Morning Post published a long-form investigation into a Chinese AI policing stack designed to infer not just identity and location, but the physical, psychological and emotional states of the people walking in front of its cameras. The system, as described by SCMP, layers facial micro-expression analysis, vocal stress detection, gait analysis and physiological proxies (heart-rate estimation from optical sensors, skin-conductance surrogates from thermal imaging) onto the existing lattice of public and commercial surveillance feeds. It is the natural next step of an architecture that has spent the last decade getting good at who you are. Now it wants to know what you are feeling, and what you might do next.
The development matters beyond China because the export market for these capabilities is already mature. City-to-city partnerships, public-safety procurement contracts and the steady rebranding of domestic policing tools as "smart city" offerings have made the underlying models a globally available commodity. The question is no longer whether state-of-the-art affect-detection will be deployed in public space, but on whose terms, under whose oversight, and against which populations the false-positive cost is permitted to climb.
What the system actually does
The SCMP account draws a careful line between the modules that have been demonstrated in published Chinese research and the modules that have been rolled into fielded products. The first category is well-established: convolutional networks trained on large internal face datasets to classify micro-expressions into the canonical emotion buckets, transformer-based voice-stress classifiers trained on call-centre and 911-style corpora, and pose/gait embeddings that hold up across re-identification checkpoints. The second category — and the one that journalists and outside researchers have had the hardest time auditing — is the fusion layer that combines these signals into a single risk score, often presented to an operator in real time on a tablet at a checkpoint or a vehicle dashboard.
Two technical claims in the SCMP report warrant a second look. The first is that pulse and respiration can be recovered from ordinary RGB video at distance, in daylight, with consumer-grade cameras. Published work over the last five years has shown this is technically possible within bounds, but the error rate under uncontrolled outdoor conditions remains a live research question, and the gap between conference demo and field deployment is exactly where civil-liberties questions tend to be decided. The second is the claim that vocal stress is a reliable indicator of deception or intent. The peer-reviewed literature is, charitably, split. The Chinese vendors cited in the SCMP piece frame these systems as "assistive" — flagging a person for human review rather than auto-action — but the documented pattern, in similar toolkits elsewhere, is that "assistive" tools have a habit of becoming decisional once the operator workload becomes a budget line.
The Chinese position, stated in its own terms
A story like this one is read very differently in Beijing than in Washington or Brussels, and the framing matters. Chinese domestic-security vendors and their state-aligned customers argue, with some structural force, that they are responding to a problem the West is also responding to: the operational gap between digital footprints and physical intent. A fraud-detection system in a Hangzhou bank, a customs-flagging system at a Shenzhen port, and a crowd-density system at a Beijing subway station are all, in vendor language, trying to do the same thing as a TSA behaviour-detection officer at a US airport — close the gap between what a person says and what a person is about to do.
The Chinese rebuttal to the dominant Western framing runs like this. First, public-safety procurement is a sovereign choice and should not be a venue for export-control theatre. Second, the underlying machine-learning capabilities are dual-use, general-purpose and already diffused across global research, and singling out Chinese vendors for restriction (as the US has periodically moved to do) risks fragmenting a market that Western firms also want to sell into. Third, the comparison class is not a hypothetical Western ideal of zero-policing surveillance, but the existing Western deployment of predictive-policing, gang-database, gang-flagging, call-pattern-analysis and acoustic-shot-detection systems — many of which have been the subject of successful civil-rights litigation at home. The structural point is hard to dismiss: there is no jurisdiction with this much camera density, this much compute at the edge, and this much appetite for fusion, that has a clean record on the question of how affect-detection is allowed to be used against its own citizens.
The counter-narrative: the same week, a different China
The juxtaposition in the SCMP news flow is the part the editorial desks in Singapore, London and New York will not want to sit with. In the same 24-hour window, SCMP also published the story of a 25-year-old Chinese man, the size of a small child, whose younger sister has gone viral online for documenting the daily caregiving routine that keeps him alive. The story is a corrective: not every Chinese household that intersects with the state does so through a camera, and not every story about dependence in China is a story about discipline. The two pieces sit next to each other in a way that exposes the fault line in Western coverage, which tends to flatten China into a single object of analysis. The family in the caregiving story is, in plain terms, a private network of love and obligation operating largely outside the formal welfare state. The subject of the policing story is, in plain terms, a public architecture designed to make every person legible to the state. Both are true. The shorthand mistake is to read one as the other.
Structural frame, in plain language
What is happening is a re-allocation of authority over a specific category of information. For most of the post-war period, the state could observe who you are and where you were, but not, in real time, what you were feeling. The closure of that gap is a civil-liberties story in every jurisdiction, but it lands first and hardest in the jurisdictions that have built the densest camera networks and the most permissive procurement environments. The export market will carry the same logic outward, packaged as efficiency, anti-fraud, anti-terrorism and public-health compliance, and the buyers will be municipal police forces and transit authorities who do not have to defend their choices in a free press. The public-interest question is not whether affect-detection works. It is who is allowed to turn it on, who is in front of it, who reviews the errors, and what the off-ramp looks like for the person on the wrong end of the score.
Stakes and the road ahead
The trajectory, if it continues, rewards the jurisdictions that buy the most cameras and the fewest lawyers, and it taxes the jurisdictions that buy the fewest cameras and the most lawyers. The Chinese vendors will continue to ship to the cities that are not subject to European-style data-protection law, and the model architectures will continue to converge, so that within a decade the difference between a Chinese system and a French system will be the procurement label, not the underlying capability. The honest version of the Western counter-argument is that an export-control regime on the model weights and training data would slow this diffusion, and the honest version of the Chinese counter-argument is that Western firms want the same contracts and have so far declined to refuse the work.
What remains uncertain, and the SCMP piece is careful about this, is the operational reliability of the fusion layer under field conditions, the audit trail available to defence counsel in any system that uses these scores, and the degree to which "assistive" scoring has already migrated into decisional workflows inside the procurers that have deployed it. The sources do not give a clean answer on any of these. The week ended with a sharper question than the one it started with: not whether China can read a face in a crowd, but whether the rest of the world is prepared to admit that it can, and that its own vendors want the same capability under a different name.
Desk note: Monexus pairs the Chinese vendor framing and the SCMP reporting without weighting either as the default; the caregiving story sits in the same news cycle as the policing story, and the difference between the two is part of the story, not a distraction from it.